The AI arms race has been quite fascinating lately. Most folks are unaware but damn, it’s all terrifying and fascinating at the same time.
LLaMa 2 is exactly what I was hoping for, and I’m glad it’s a thing now. At first glance, it’s not as great as GPT-3.5/4 or PaLM but for now, I’m assuming it’s the lack of my own hardware and fine-tuned models. That said, it is still very impressive.
GPT-3.5/4 are powerful, but I have to admit, I’m seeing the nerfs now and it is a bit distressing to see its potential taking a few steps back while it also improves with the new betas they’ve opened up. I’ve still yet to use the code interpreter stuff, for example.
I’m so glad GGML is a thing. I’ve heavily used Whisper (speech transcription from OpenAI) to great effect for several months now and it’s all properly usable thanks to this project.
Finally, Stable Diffusion is an entire rabbit hole on its own too. There’s crazy stuff there I haven’t explored, and there’s a reason why even Photoshop now uses similar tech too.
My MacBook Pro is now finally being put to the test with these AI stuff. It’s always been the case with SD, but with LLMs, the lack of RAM is really showing (8GB is unusable, and 16GB is passable but still low, seriously! The larger models require around 60GB, not kidding).
The community’s very active at optimizing stuff though, and I want to see the fancy Neural Engine to actually be properly used someday. It’s really improved very well though, from raw painful Python to GGML C++ runs with experimental Metal support.